What I’m working on at the moment. Updated monthly — a now page, not a changelog.
Building
k8s-gpu-scheduler-lab, phase 1 of 3,
now public. The kwok substrate, the workload generator, the metrics and the
degenerate baselines are built, and K0 — the default kube-scheduler — is
measured on a real control plane. Phase 2 is Kueue and Volcano, which is where
the comparison actually starts. The 34% bin-packing claim the repo exists to
test has not been tested yet, and its row stays empty until it has.
Also shipped this month: slurm-mcp and cluster-ops-skills — the read-only tool surface an agent reads a cluster through, and the runbooks it follows once it can.
Still in flight: cluster-sre-agent, phase 3 of 6. The dependency graph and read-only tool surface are built and tested; the LLM configurations need an API key and are next. The graph is the interesting part — 62% of its edges are measured on a live cluster rather than assumed.
Writing
Next post: what happened when I tried to reproduce the storage-stall-to- scheduling-halt chain everyone repeats, and could not. It is the most useful negative result I have measured this year.
Submitting
Talks and papers in flight, so there is one place to look rather than an announcement after the fact.
Nothing under review right now. Two things are close enough to name:
- The Kubernetes GPU scheduler comparison, once phase 2 has Kueue and Volcano numbers against the same trace. A controlled cross-scheduler comparison is the kind of thing SUG or KubeCon exists for, and it is not worth submitting before there are real numbers in it.
- The storage-stall refutation — a negative result about a widely-repeated causal chain, which is a better lightning talk than a paper.
This section stays honest about the difference between submitted, accepted and thinking about it. Anything listed here without a status is in the last category.
Learning
MS CS at Georgia Tech — machine learning specialisation. Alongside it, reading Slurm’s scheduler source rather than only its documentation, which is where the difference between “documented” and “actually true” keeps turning up.
Reading
Advances in Financial Machine Learning (López de Prado) — chapter 7 on temporal separation and embargoes is doing real work in the evaluation harness for research-platform.
Developing Time-Oriented Database Applications in SQL (Snodgrass) — the standard treatment of bitemporal modelling, and the reason the point-in-time store looks the way it does.
Last updated: 23 August 2026.